Developer Experience in the Age of AI Coding Agents (Category Agents)
I watched an interesting talk by Max Kanat-Alexander, Executive Distinguished Engineer at Capital One, about DevEx in the age of AI agents—or how to avoid drowning in existing legacy code and the new legacy code agents are producing at speed :) Max previously worked as a Tech Lead at Google on Code Health and as a Principal Staff Engineer at LinkedIn on Developer Productivity. He also wrote “Code Simplicity” and “Understanding Software”, which I have not read yet :)
The main points of the talk:
1️⃣ Don't fight the training set Use standard tools. If you have written your own package manager or use an obscure language, the agent will struggle. Its training is grounded in common open-source tools. The more conventional and “boring” your stack, the better AI can work with it. 2️⃣ CLI > GUI Agents need APIs and CLIs rather than a browser. Making an agent click through a GUI is expensive and unreliable. A text interface lets it work faster and more accurately. 3️⃣ Tests must be deterministic An error such as `500 Internal Error` tells an agent very little. It needs clear validation messages. Investing in useful test and linter errors is an investment in agent autonomy. Tests also need to run quickly—30 seconds rather than 20 minutes. The agent runs them repeatedly, so slow CI undermines its productivity. 4️⃣ Document why, not just what An agent can see the code and understand what it does. It was not in your meetings, however, and cannot read minds. Documentation needs context: business goals, external constraints, and the shape of input data—information that is not in the code. 5️⃣ The code-review problem: a vicious cycle Writing code is becoming a matter of reading it, while the number of pull requests grows exponentially. With a weak review process, teams start rubber-stamping poor agent-written code with “LGTM.” The codebase deteriorates, making it harder for agents to work in, and they produce even worse code. The response is to distribute review responsibilities, rather than send everything to a general “could someone look at this?” channel, and maintain a high quality bar.
🚀 What does this mean for development? - Legacy refactoring is essential. If a person cannot understand a project's structure without undocumented knowledge, an agent will hallucinate its way through it. Good code structure is now an economic necessity. - A shift toward verification. We are moving from writing code to checking it. Being able to read and validate someone else's code quickly is becoming more important than typing quickly. - The golden rule. Everything that helps an AI agent—fast tests, clear errors, standard tools—also helps people. Even if AI disappeared tomorrow, these investments would still benefit human developers. Amusingly, this resembles the “Golden Rule”: treat others as you would want them to treat you.
#Engineering #AI #Metrics #Software #DevEx #Productivity #DevOps #Architecture #Culture #Engineering #ML #SystemDesign